Upload a simulated or research-use GuiTu XML-format 12-lead ECG file. ECG-ArrestNet extracts electrophysiologic features, runs the full inference pipeline, and returns a calibrated 24-hour IHCA probability with explainable contributions.
ECG-ArrestNet is a research prototype developed to demonstrate an ECG-based deep learning workflow for estimating short-term IHCA risk after index ECG acquisition.
The Second Xiangya Hospital of Central South University served as the development center and provided the development and temporal-validation cohorts. Independent external validation was performed at Xiangya Hospital of Central South University and The Central Hospital of Enshi Tujia and Miao Autonomous Prefecture.
Multi-scale 1D-CNN backbone, 8-head lead-aware cross-attention, BiLSTM temporal encoder, and gated fusion of 52 ECG-derived features with deep-learning representations.
Uploaded demonstration ECG files are processed locally for the current session and are not retained by the platform.
Upload a simulated or research-use GuiTu XML-format 12-lead ECG file, then run the prediction. The model uses information derived exclusively from the uploaded 12-lead ECG; demographic and non-ECG clinical variables are not model inputs. This in-browser demonstration reproduces the ECG-ArrestNet inference workflow for research illustration; the output includes a calibrated 24-hour IHCA probability, a descriptive risk stratum (Low <1%; Intermediate 1–<5%; High ≥5%), feature-level contributions, and a lead-aware attention visualization.
This platform is a research demonstration of the ECG-ArrestNet inference workflow. It is not a medical device and must not be used for diagnosis, triage, monitoring, or treatment decisions. As stated in the study, the model was evaluated retrospectively; prospective validation and subsequent clinical-impact evaluation are required before clinical adoption. The demonstration ECG files are simulated materials used solely for interface and workflow testing and were not used for model development, calibration, temporal validation, external validation, or performance estimation. Uploaded signals are processed locally in the browser and are not retained by the platform.
Reported performance was derived from retrospective inpatient cohorts assembled using prespecified eligibility criteria without case-control sampling or outcome-based enrichment.
Model probabilities were calibrated using the held-out internal-validation set. The calibration mapping was subsequently applied without modification to the temporal-validation and external-validation cohorts, with no center-specific recalibration.